July 17th, 2013

Learning Auxiliary Dictionaries for  Undersampled Face Recognition Chia‐Po Wei and Yu‐Chiang Frank Wang Research Center for IT Innovation, Academia Sinica, Taipei, Taiwan

IEEE International Conference on Multimedia & Expo 2013

Outline • Introduction  Undersampled Face Recognition • Related Work  Sparse Representation based Classification (SRC) & ESRC  Dictionary Learning for Sparse Representation • Learning of Auxiliary Dictionary • Experiments • Conclusions

2

Why Face Recognition? Computational Forensics 

Access Control

Mass Surveillance

People Identification/Search

3

Challenges • Robust Face Recognition  Face images with possible illumination and expression variations   Face images might be corrupted due to occlusion or disguise.  A sufficient amount of training data is required for handling such variations.

• Undersampled Face Recognition  Practically, only one or very few images per subject are available for training.  It would be very difficult to handle both intra and inter‐class variations.

4

Approaches to Undersampled FR Problems • AGL (Su et al., CVPR’10): Adaptive generic learning • ESRC (Deng et al., PAMI’12): Extended sparse representation‐ based classification • PCRC (Zhu et al., ECCV’12): Multi‐scale patch based  collaborative representation approach • Remarks  AGL and PCRC did not considered corrupted face images .  ESRC directly applied external data for modeling intra‐class variations.  How to properly utilize external data remains a question…

5

Outline • Introduction  Undersampled Face Recognition • Related Work  Sparse Representation based Classification (SRC) & ESRC  Dictionary Learning for Sparse Representation • Learning of Auxiliary Dictionary • Experiments • Conclusions

6

Sparse Representation based Classification (SRC) • •

Consider an input image as a sparse linear combination of training images. Given a test image y and training data D = [D1, D2, …, DC] of C classes,  SRC (Wright et al., PAMI’09) derives the sparse coefficient x of y as:

min y  Dx 2 + x 1 2

x



The identity of y will be given by

arg min y  D k (x) k



2 2

y = [D1, D2, …, DC] x + ε =       D x + ε

i.e., classification is based on class‐wise minimum reconstruction error. However, SRC requires a large amount of training data as D.

7

Extended SRC •

To improve SRC, Deng et al. (PAMI’12) proposed  extended SRC (ESRC): 2

x  min y  [D, A ]  a  + x 1 , where x  [x d ; x a ] x x  2 d



Intra‐class variant dictionary A  Contains images collected from an external dataset (subjects not of interest)  ESRC performs classification by

 k (x )  arg min y  [D, A ]  a  k  x  d



2

2

ESRC utilizes external data to model intra‐class variations.  Performance might degrade due to noise or redundant images presented.  The size of the external dataset can be very large (and thus a large A). 8

Outline • Introduction  Undersampled Face Recognition • Related Work  Sparse Representation‐based Classification (SRC) & ESRC  Dictionary Learning for Sparse Representation • Learning Auxiliary Dictionaries • Experimental Results  Extended Yale B  AR • Conclusions 9

Dictionary Learning (DL) for Sparse Representation • Unsupervised Dictionary Learning: Data Representation  KSVD (Aharon et al., TIP’06)  MOD (Engan et al., ICASSP’99) 

• Supervised Dictionary Learning: Data Separation    

Supervised Dictionary Learning (Mairal et al., NIPS’09) Discriminative K‐SVD (Zhang and Li, CVPR’10) Label Consistent K‐SVD (Jiang et al., CVPR’11) Fisher Discrimination Dictionary Learning (Yang et al., ICCV’11)

• Remarks  The above approaches might not generalize well if limited training data. (i.e., for undersampled recognition problems)  For undersamepled FR, possible integration of ESRC and DL? 10

Outline • Introduction  Undersampled Face Recognition • Related Work  Sparse Representation based Classification (SRC) & ESRC  Dictionary Learning for Sparse Representation • Learning of Auxiliary Dictionary • Experiments • Conclusions

11

Our Proposed Method x

D

A

y

2

x  min y  [D, A ]  a    x 1 x x  2 d

 k ( x )  identity  min y  [D, A ]  a  k  x  d

2

2 12

Our Contributions • Learning of Auxiliary Dictionary A  Observing intra‐class variations using external data

• Advantages:  Only one or few training images per subject class required.  The auxiliary dictionary A is derived from external data (i.e., subjects  NOT of interest).  A compact A is observed: preferable for storage/computation  Can deal with corrupted face images in training and testing

13

Problem Formulation • Our Proposed Algorithm (for learning A) N

min  y  [D , A]xi A ,X

i 1

e i

e

Reconstruction error

2 2

  xi 1   y  D  ik (x )  Ax

Sparsity constraint

e i

e

d i

2 a i 2

Classwise reconstruction error  for class k

e e e e  Y  [ y1 , y 2 , , y N ] : probe set of external data

 De : gallery set of external data  A : auxiliary dictionary to be learned  xi  [xid ; xia ]: sparse coefficient of  y ie

• Remarks  Need to learn the auxiliary dictionary A and sparse coefficients x during training using external data. 14

The Training Stage • Optimization via iterating between sparse coding and  dictionary updates • Sparse Coding for Updating X  Given fixed A, we rewrite our formulation as: N

min  X

i 1

2

A  x  y   D  e    (De )  A      xi   x  2  y i   ik e i

e

d i a i

1

where    1/2 , and  i (De )  [0,0, , Die , ,0] k k  The above is the standard L1‐minimization problem.  Can be solved via techniques like Homotopy, Iterative Shrinkage‐ Thresholding, or Augmented Lagrange Multiplier (ALM) 

15

The Training Stage (cont’d) • Dictionary Update for A:  We fix X and rewrite the objective function as N

min  A

i 1

 y D x   Ax   e  e d  y D x     ( ) Ax   i ik i     e i

e

d i

a i a i

2

2

 Setting the partial derivatives wrt aj (the jth column of A) equal to zero,  the analytically solution can be derived as

A  V T (U 1 )T , where  xi1yˆ Ti Z  e e d N N   y D xi  I   i 2 a a T , U  (1   ) xi (xi ) , V     , yˆ i   e Z  e d   I  ( ) y D x     i 1 i 1   i i i   T k    xim yˆ i Z   

where

16

The Testing Stage • Inputs:  Query y  Undersampled training data D (collected from the subjects of interest)  Auxiliary dictionary A (learned from external data)

• We solve

2

x  min y  [D, A ]  a    x 1 x x  2 d

• Once x is obtained, recognition is achieved by  k (x d )  identity  min y  [D, A ]  a  k  x 

2

2

17

Remarks • Recall that our proposed formulation solves: N

min  y  [D , A]xi A ,X

e i

i 1

e

Reconstruction error

2 2

  xi 1   y  D  ik (x )  Ax e i

Sparsity constraint

e

d i

2 a i 2

Classwise reconstruction error for class k

• Comparisons with SRC‐based Approaches: Corrupted  Training Data

Undersampled Training/Gallery Set

Dictionary  Learning

SRC

X

X

X

LR

O

X

X

ESRC

O

O

X

Ours

O

O

O

 [SRC] Wright et al., Sparse representation‐based classification, PAMI’09  [LR] Chen et al., Low‐rank matrix recovery with structural incoherence, CVPR’12  [ESRC] Deng et al., Extended SRC, PAMI’12 18

Outline • Introduction  Undersampled Face Recognition • Related Work  Sparse Representation based Classification (SRC) & ESRC  Dictionary Learning for Sparse Representation • Learning of Auxiliary Dictionary • Experiments • Conclusions

19

Extended Yale B • 38 subjects available, each has about 64 images • External data for training the auxiliary dictionary:  6 subjects (out of 32 considered not of interest), 64 images each

• Gallery set:  32 subjects (of interest), each has 3 images considered available

• Test set:  32 subjects, each has about 61 images for evaluation Training Images

Test Images

20

Comparisons with ESRC using Different Features

 ESRC randomly selects elements from external data as A. 21

Performance Comparisons  with ESRC, K‐SVD, SRC, and NN LBP Features

Pixel‐based Features

 For KSVD, ESRC, and our approach, the number of external subjects is set as 2,  while NN and SRC does not utilize any external data. 22

AR Database • External Data (for training the auxiliary dictionary) 20 subjects (out of 100 considered not of interest), each has 13 images

• Gallery Set 80 subjects; only 1 neutral image per subject is available.

• Test Set 80 subjects, each has about 12 images Single Gallery Image

Test  Images

23

Performance Comparisons  with Different Auxiliary Dictionary Sizes

 Note that PCRC and SRC do not utilize external data for recognition,  while AGL was not designed for handling occluded face images. 24

Performance Comparisons with Different Feature Dimensions LBP Features

Pixel‐based Features

 For KSVD, ESRC, and our approach, the number of external subjects is set as 2,  while NN and SRC does not utilize any external data. 25

Outline • Introduction  Undersampled Face Recognition • Related Work  Sparse Representation based Classification (SRC) & ESRC  Dictionary Learning for Sparse Representation • Learning of Auxiliary Dictionary • Experiments • Conclusions

26

Conclusions • We propose to learn an auxiliary dictionary from external data for undersampled face recognition. • The auxiliary dictionary is derived for handling intra‐class  variations, which is more compact and representative than  ESRC, etc. approaches. • Together with ESRC‐based formulations, our proposed method  achieved improved recognition results than state‐of‐the‐art  SRC‐based approaches.

27

Thank You!

28

Learning Auxiliary Dictionaries for Undersampled Face ...

Jul 17, 2013 - dictionary updates. • Sparse Coding for Updating X. ➢ Given fixed A, we rewrite our formulation as: where. ➢ The above is the standard L1-minimization problem. ➢ Can be solved via techniques like Homotopy, Iterative Shrinkage-. Thresholding, or Augmented Lagrange Multiplier (ALM). 15. 2. 1. 1. 2 min.

875KB Sizes 0 Downloads 153 Views

Recommend Documents

Undersampled Face Recognition via Robust Auxiliary ...
the latter advanced weighted plurality voting [5] or margin ... Illustration of our proposed method for undersampled face recognition, in which the gallery set only contains one or ..... existing techniques such as Homotopy, Iterative Shrinkage-.

Jointly Learning Visually Correlated Dictionaries for ...
washer; grand piano; laptop; pool table; sewing .... (9). The matrix Mi. ∈ RKi×Ni consists of Ni copies of the mean vector µi as its columns. And the matrices. M0.

Probabilistic learning for fully automatic face ...
We propose novel extensions by introducing to use a more robust feature description as opposed to pixel- based appearances. Using such features we put forward ..... Thirteen poses covering left profile (9), frontal (1) to right profile (5), and sligh

Learning Prototype Hyperplanes for Face Verification in the Wild
(referred to as the prototype hyperplane) of one Support Vector. Machine (SVM) model, in which a sparse set of support vectors is selected from the unlabeled ...

LNCS 4233 - Fast Learning for Statistical Face Detection - Springer Link
Department of Computer Science and Engineering, Shanghai Jiao Tong University,. 1954 Hua Shan Road, Shanghai ... SNoW (sparse network of winnows) face detection system by Yang et al. [20] is a sparse network of linear ..... International Journal of C

Auxiliary Parameter MCMC for Exponential ... - Semantic Scholar
Keywords ERGMs; Parameter inference; MCMC; Social Networks; ... sociology [4], political sciences [5], international relations [6], medicine [7], and public health ...

Auxiliary Membership.pdf
Page 1 of 1. American Legion Auxiliary. World's largest women's patriotic service organization. MEMBERSHIP APPLICATION. APPLICANT INFORMATION.

Business Auxiliary Services.pdf
as amended by Clause (f) of Section 107 and Clause (2) of Section 109 of. Finance Act, 2015 ... legislative competence to enact the said amendments since the activity of ... under Entry 51 of List II of Schedule VII to the Constitution. The case of .

adjectives in dictionaries
from the Bank of English corpus for each adjective are analysed before their respective dictionary entries are evaluated. 2 ...... hours …singing the eight canonical hours./ …the prayers were arranged according to the canonical hours of the day./

Extraction Of Head And Face Boundaries For Face Detection ieee.pdf
Loading… Whoops! There was a problem loading more pages. Whoops! There was a problem previewing this document. Retrying... Download. Connect more apps... Extraction Of ... ction ieee.pdf. Extraction Of H ... ection ieee.pdf. Open. Extract. Open wit

Extraction Of Head And Face Boundaries For Face Detection.pdf ...
Extraction Of Head And Face Boundaries For Face Detection.pdf. Extraction Of Head And Face Boundaries For Face Detection.pdf. Open. Extract. Open with.

Auxiliary Power Unit.pdf
Retrying... Download. Connect more apps... Try one of the apps below to open or edit this item. Auxiliary Power Unit.pdf. Auxiliary Power Unit.pdf. Open. Extract.

Subject-Auxiliary Inversion
Jun 26, 2013 - The treatment of English auxiliary verbs epitomizes the stark contrast between the original generative view of syntax as an autonomous formal ...

Auxiliary Parameter MCMC for Exponential Random Graph Models*
Abstract. Exponential random graph models (ERGMs) are a well-established family of statistical models for analyzing social networks. Computational complexity ...

auxiliary verb examples pdf
Sign in. Loading… Whoops! There was a problem loading more pages. Whoops! There was a problem previewing this document. Retrying... Download. Connect ...

Diesel Reforming for Fuel Cell Auxiliary Power Units
Jan 27, 2005 - Carbon Heat Capacity Determination. Heat capacities for C. 2. * amorphous carbon from. Gibb's Free Energy data. -14000. -12000. -10000.

Topology and Control Innovation for Auxiliary ... - Research at Google
provide a constant current (CC) driving current preferably for the best light output ... As the Internet of Things (IoT) continues to proliferate, connected and smart .... communicate (via wireless network) and interact with the environment (via ...

Topology and Control Innovation for Auxiliary ... - Semantic Scholar
LED lighting applications will require forced-air convection because of the high intensity of light needed such as recessed light, spot light and track light. Regular fans, however, introduce life time and reliability concern and for indoor applicati